`VirtualGraphComposite` underperforming on current QPUs

[JoelPasvolsky]

Description I don't have saved data from previous QPU generations for comparison but for current QPUs (with their improved calibration) VirtualGraphComposite is likely undersampling magnetization measurements. Below are resulting FBOs for several executions 5 seconds apart on the Advantage_system4.1 QPU:

sampler_vg = VirtualGraphComposite(qpu, {'a': [3000, 300, 2971, 345, 3136, 240]}, 2, flux_bias_max_age=5)
>>> sampler_vg.flux_biases[300]
7.928525473822266e-06
-2.499999999999999e-05
-1.996443354549972e-09
0.0
1.0130290919593328e-09
-2.4999999999999998e-05
2.499999999994167e-05

Notice especially the last two results.

To Reproduce See above

Expected behavior More or less consistent FBO for the chain.

Environment:

• OS: WIN10
• Python version: 3.10

Additional context Consider obsoleting the composite versus updating the algorithm, from some recent testing it seems to me that for current QPUs one needs tens of thousands of samples for chain magnetization to be meaningful, so the benefit is marginal at best.